Do people have to learn from dogs when doing AI artificial intelligence? researchers say yes

What can artificial intelligence learn from dogs?

This is obviously a controversial but interesting topic. We have seen a lot of interesting research before, such as pig face recognition, strawberry recognition in the plant kingdom, and more classic cat face recognition.

At present, a group of elite researchers seem to have already shared some results from the study of the relationship between dogs and artificial intelligence. One of the important conclusions is that it is time-consuming to program algorithms through many rules, but dogs already know everything. According to the rules, is the robot dog the best robot? ?

Enjoy the following:

Researchers at the University of Washington (University of Washington) and the Allen Institute for Artificial Intelligence (Allen Institute for AI) say that there are already research results in this area.

They recently trained a neural network to interpret and predict canine behavior. Their results suggest that animals could provide a new source of training data for artificial intelligence systems, including those used to control robots.

To train an AI to think like a dog, researchers first need data. They collected this information in the form of video and motion information captured from a Malamute named Kelp.

The researchers captured a total of 380 videos from a GoPro camera mounted on the dog's head, as well as motion data from sensors on the dog's legs and body. Essentially, "Kelp" was recorded in the same way Hollywood uses motion capture to record CGI-created actors, capturing a dog going about its daily activities, such as walking, playing fetch and going to the park.

With this information, the researchers analyzed Kelp's behavior using deep learning methods. It's an artificial intelligence technique that can be used to sift through data for patterns.

In this case, that meant matching Kelp's limb movement data with the GoPro's vision data to various dog activities. The resulting neural network was able to predict what the dog would do in certain situations. For example, if it sees someone throw a ball, then it knows the dog's response is to turn around and chase it.

The paper's lead author, Kiana Ehsani, explained in an interview with The Verge that their AI system's predictive power was very accurate, but only for a short period of time. For example, if the video shows a set of stairs, you can guess that the dog will climb them.

But based on the diversity of real life, there are still uncertainties. "Who knows if the dog will see a toy or something it wants to chase," said Ehsani, a doctoral student at the University of Washington.

However, what the researchers did next showed its clever features. With neural networks trained on dog behavior, they wondered whether dogs learned other things about the world that they hadn't been explicitly programmed to do.

As they explain in their paper, dogs "clearly demonstrate visual intelligence, recognizing food, obstacles, other people, and animals," so could a neural network trained to act like a dog display the same intelligence? ?

As it turns out, the answer is yes, though only in very limited circumstances. The researchers conducted two tests on the neural network, asking it to recognize different scenes (such as indoors, outdoors, stairs, balconies) and "walkable surfaces" (sounds exactly the same: places you can walk on). In both cases, neural networks are able to perform these tasks with very high accuracy using only basic data.

"Our intuition for this is that dogs are really good at finding places to walk, where they can go, where they can't go," Ehsani said. "That's a very difficult task for a computer because It requires a lot of prior knowledge." That knowledge might be whether a surface is too steep to walk on, or too pointed and uncomfortable.

Programming a robot with all these rules is time-consuming, but dogs already know all the rules. Thus, by observing Kelp's behavior, the neural network learns these rules without learning. In other words, it learned from the dog.

Now, it's important to include a lot of caveats here. The software Ehsani and her colleagues developed is by no means a model of a dog's brain or its consciousness. All it does is learn from a limited set of data some very basic rules of where dogs like to walk.

As with other AI systems, there is no reasoning here, the software just looks for patterns in the data, which itself is not new, researchers train AI systems from similar data all the time.

But, as Ehsani points out, this appears to be the first attempt by anyone to learn from dogs, and its work suggests that animals can be a useful source of training data.

After all, dogs know a lot of data that would be useful for robots, among other things. Dogs know how to avoid cars and how to navigate stairs, important lessons for any robot that needs to operate in a human environment.

Of course, this article is just a simple demonstration of what we can learn from animals, and more work is needed before this paradigm can be fruitful. But Ehsani says she's confident that some very useful applications can be drawn from it. "One of the things that came to my mind was to make a robot dog.

"It's already a difficult task for a robot to know how to move and where to go, or what they want to chase. This will definitely help us make more efficient and better robot dogs," she said.